LightTrack: Finding Lightweight Neural Networks for Object Tracking via One-Shot Architecture Search

Environment Installation

cd lighttrack
conda create -n LightTrack python=3.6
conda activate LightTrack
bash install_env.sh

Data Preparation

• ImageNet

Please first download the ImageNet-2012 then unzip it to the folder $LightTrack/supernet_backbone/data/imagenet and move the validation set to the subfolder $LightTrack/supernet_backbone/data/imagenet/val. To move the validation set, you cloud use the following script: https://raw.githubusercontent.com/soumith/imagenetloader.torch/master/valprep.sh

The prepared ImageNet dataset should look like:

$LightTrack/supernet_backbone/data/imagenet/train
$LightTrack/supernet_backbone/data/imagenet/val
...

• Tracking Training Set

Please download training data from GoogleDrive, and then unzip them into $LightTrack/data. The prepared tracking training set should look like:

$LightTrack/data/got10k
$LightTrack/data/y2b
$LightTrack/data/vid
$LightTrack/data/det
$LightTrack/data/coco

• Tracking Validation Set (for evolutionary search)

In this work, we choose GOT-10K val set as the tracking validation set. Please download it from GOT10K-Val then unzip it into $LightTrack/GOT-10K. The prepared tracking validation set should look like:

$LightTrack/GOT-10K/val/GOT-10K_Val_000001
$LightTrack/GOT-10K/val/GOT-10K_Val_000002
...

• Tracking Benchmarks

Please put VOT2019 dataset under $LightTrack/dataset. The prepared data should look like:

$LighTrack/dataset/VOT2019.json
$LighTrack/dataset/VOT2019/agility
$LighTrack/dataset/VOT2019/ants1
...
$LighTrack/dataset/VOT2019/list.txt

LightTrack Pipeline

• Step1: Pretraining Backbone Supernet

cd supernet_backbone
chmod +x tools/distributed_supernet.sh
bash experiments/supernet_pretrain.sh
cd ..

• Step2: Training Tracking Supernet

python tracking/onekey_X_supernet_simple.py --cfg experiments/train/supernet_train.yaml --WORK_DIR . --back_super_dir supernet_backbone/experiments/search/hypernet

• Step3: Searching with Evolutionary Algorithm on Tracking Supernet

bash Evolution/src/Search/search.sh

Flops, Params, and Speed

Compute the flops and params of our LightTrack-Mobile. The flops counter we use is pytorch-OpCounter

python FLOPs_Params.py

Test the running speed of our LightTrack-Mobile

python Speed.py

Test and evaluation

Test LightTrack-Mobile on VOT2019

bash reproduce_vot2019.sh